This thesis explores advanced methods in causal inference, focusing on policy learning, instrumental variables (IV), and difference-in-differences (DiD). First, it addresses challenges of restrictive assumptions in IV and DiD methods, proposing an instrumented DiD approach to relax them. The thesis also introduces direct policy search to learn optimal policies, which come with novel identification results, and various estimators, ensuring robust policy learning under unmeasured confounding. It also presents a positivity-free policy learning framework using dynamic policies, enhancing incremental intervention effects with efficient machine learning methods. Additionally, it proposes a transfer learning framework for individualized treatment regimes in heterogeneous populations with survival data, offering robust tools for practical applications.

Page web: https://team.inria.fr/premedical/phd-defense-pan-zhao/